Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Long-term retrospective medical imaging archives cannot be directly utilized due to the lack of fine-grained annotations, requiring extensive manual curation by expert physicians in most previous medical imaging studies. Furthermore, the absence of standardized archiving protocols and the heavy reliance on subjective judgment for image retention lead to significant variations in the number of images per patient, while the presence of irrelevant frames introduces substantial noise. We propose FARSIGHT, an end-to-end archive-to-prediction framework designed to overcome these universal bottlenecks, evaluated here on renal transplant chronic allograft injury assessment. FARSIGHT integrates the Modality Aware Governance and Optimization System (MAGOS), an offline vision-language-model-guided curation agent, with hierarchical multi-modal multi-instance learning (HMMIL) for patient-level prediction. On a retrospective cohort of 4,576 patients comprising 41,203 images across four distinct modalities, MAGOS achieves 0.999 routing accuracy with complete parsing success. Using these curated inputs, HMMIL achieves a test AUC of 0.816 and an accuracy of 0.746 for patient-level prediction, while handling variable-length studies and missing modalities. The code is available at: https://github.com/hanlliumu/Farsight.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0806_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/hanlliumu/Farsight

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YanJin_FARSIGHT_MICCAI2026,
        author = { Yang, Jinghan AND Li, Qinghua AND Liu, Jingxin AND Qiu, Li AND Tian, Jie AND Zhou, Jiaojiao AND Wang, Kun},
        title = { { FARSIGHT: A Dynamic Multi Modal Medical Image Analysis Framework Powered by a Prior Knowledge Guided Vision Language Model } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    FARSIGHT is an end-to-end archive-to-prediction framework addressing universal bottlenecks in real-world retrospective medical imaging archives. The main contributions are: 1.Magos, an offline VLM-guided curation agent that performs modality routing via prior-knowledge-encoded visual state space and deterministic rule mapping, combined with spatial heuristic adaptive ROI extraction — achieving 99.98% routing accuracy on 40,000+ images without foundation-model fine-tuning. 2.HMMIL, a hierarchical multi-modal multi-instance learning framework featuring two-level attention aggregation, masked two-stage Transformer fusion, and auxiliary Top-K hard negative instance supervision — enabling patient-level prediction under variable-length studies and missing modalities. 3.Demonstrated on renal transplant chronic allograft injury assessment, HMMIL improves test AUC from 0.777 to 0.816 over the strongest baseline, validating robustness under real-world curation noise and missing-modality conditions.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Clinically Motivated Problem with Broad Relevance The paper targets universal bottlenecks in real-world retrospective archives — “lack of fine-grained annotations, unreliable metadata” and studies with “variable image counts, redundant or noisy frames, and missing modalities” — a challenge highly representative of large-scale clinical practice. 2.Magos: Zero Fine-Tuning Curation with Auditable, Privacy-Preserving Design Magos achieves “privacy-preserving on-prem deployment” without any annotated training samples by decoupling VLM perception from deterministic rule mapping and enforcing “JSON-Schema-constrained decoding” for structured, verifiable outputs. This design is both practically innovative and deployable under strict institutional constraints. 3.Convincing Large-Scale Empirical Validation Deployed on 40,000+ uncurated images, Magos achieves “99.98% routing accuracy” with “complete parsing success rate without a single pipeline crash,” substantially outperforming a supervised ResNet18 baseline (66.7% on Power Doppler) in a zero-shot setting. 4.HMMIL Design Aligns with Clinical Acquisition Logic The two-level hierarchy — intra-modality attention followed by inter-modality Transformer fusion — explicitly mirrors clinical workflows where “instances are first filtered within modality before any cross-modality interaction,” yielding robustness under variable-length and incomplete studies, with each component validated through ablation. 5.Top-K Hard Negative Loss Addresses a Known MIL Weakness Standard MIL tends to “assign high attention to non-diagnostic frames, especially in negative patients. “ The proposed auxiliary loss directly targets this failure mode and is shown to improve both prediction performance and instance-level saliency reliability. 6.Strong Clinical Translation Potential The system “requires no changes to physician workflows” and is compatible with existing PACS infrastructure, offering a practical path to automating retrospective cohort construction at scale.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The framework is validated exclusively on one private hospital dataset for CAI assessment, with no external multi-center validation or public benchmark testing. The generalizability of Magos’s spatially tuned heuristic rules to different device manufacturers or acquisition protocols remains unconfirmed.

    2.The reported B-mode mIoU of 62% falls well below the paper’s self-defined acceptance criterion of 0.80, yet this is rationalized as “sufficient for patient-level prediction tasks” without rigorous ablation evidence. Given that B-mode is identified as “the most informative single modality,” the downstream impact of this coarse extraction warrants more careful quantitative analysis.

    3.Table 2 does not clearly specify whether modality ablation and module ablation share a consistent baseline configuration, creating ambiguity in the conclusions. Notably, B+C+S (Test AUC 0.808) outperforms other modality combinations in the modality ablation, yet the final HMMIL model uses only B+C and reports AUC 0.816.This improvement likely reflects the contribution of the full module configuration rather than the modality choice itself. Critically, the paper never reports B+C+S with full modules, making the final modality selection decision experimentally unjustified.

    4.All baselines in Table 1 are single-modality, general-purpose MIL methods. No multimodal MIL or multimodal fusion baselines are included, making it difficult to determine whether HMMIL’s gains stem from the hierarchical design or simply from incorporating multiple modalities. A more comprehensive comparison with relevant prior multimodal MIL works would be needed to substantiate the claimed contributions.

    5.All performance metrics are reported as point estimates without confidence intervals, standard deviations, or significance tests. Given the modest absolute improvements (e. g. , AUC 0.777 → 0.816), it remains unclear whether the differences are statistically significant, which does not meet the MICCAI reviewer guideline requirement for “performance measures reported with measures of uncertainty or confidence. “

    6.Hard negative mining in MIL is not new — MHIM-MIL (Tang et al. , ICCV 2023, ref. 11) already proposes masked hard instance mining for WSI classification. The paper should more clearly articulate how its asymmetric treatment of positive and negative bags differs from existing approaches to substantiate its novelty claim.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses a genuinely important and underexplored problem — automating curation and prediction for real-world retrospective medical imaging archives — and presents a complete end-to-end framework with two technically coherent components. The large-scale empirical validation of Magos is a practical strength that distinguishes this work from purely algorithmic contributions. The hierarchical MIL design is well-motivated and the ablation study, despite its limitations, provides reasonable support for the proposed components. The primary factors limiting a stronger recommendation are the inconsistent ablation controls in Table 2 — particularly the absence of a B+C+S with full modules comparison — which leaves the central modality selection decision experimentally unjustified. The lack of multimodal MIL baselines also makes it difficult to fully attribute the performance gains to the proposed hierarchical design. Additionally, the single-center validation and absence of statistical uncertainty measures reduce confidence in the reported improvements. Nevertheless, the work makes a meaningful contribution to a practical clinical challenge, demonstrates strong system-level performance, and is of clear interest to the MICCAI community working on real-world medical image analysis. These strengths are sufficient to justify acceptance, provided the authors address the ablation inconsistency and statistical reporting in their rebuttal.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I appreciate the authors’ rebuttal and acknowledge that several key concerns have been effectively addressed. Most importantly, the authors clarified that the full-module B+C+S result is already reported in Table 2, directly resolving my primary concern regarding ablation inconsistency. The explanation that S modality contributes limited diagnostic value — supported by the attention analysis — provides reasonable justification for the final B+C modality selection. The acknowledgement that 95% CIs were removed due to space constraints, with intention to restore them in consultation with the AC, is also appreciated. Remaining concerns include the lack of multimodal MIL baselines, which still makes it difficult to fully attribute performance gains to the hierarchical design rather than simply to multi-modality input, and the insufficiently supported B-mode IoU justification, as the referenced preliminary trials were not reported. Overall, the rebuttal resolves the most critical experimental concern. I am updating my score to 5 and recommend acceptance, though the absence of multimodal baselines and incomplete statistical reporting remain limitations that should be addressed in the camera-ready version.



Review #2

  • Please describe the contribution of the paper

    This paper proposes FARSIGHT, an end-to-end archive-to-prediction framework for real-world medical imaging repositories, designed to address the practical challenges of retrospective clinical data, including missing fine-grained annotations, irregular study lengths, noisy frames, and incomplete modality availability. In addition, the paper presents Magos, an offline and auditable VLM-plus-rules curation module for modality routing and ROI extraction without foundation-model fine-tuning, and HMMIL, a hierarchical multi-modal MIL framework designed for variable-length and missing-modality clinical data. Extensive experiments on the renal transplant ultrasound cohort demonstrate that the proposed framework achieves strong curation reliability and predictive performance.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.This paper integrates clinical archive data governance with patient-level prediction, effectively addressing key challenges in retrospective clinical archives, including missing fine-grained annotations, irregular numbers of images per patient, noisy frames, and missing or misaligned modalities. 2.Magos combines VLM perception with prior-knowledge rules, JSON-schema-constrained decoding, a fallback parser, and deterministic ROI extraction. The strategy makes the pipeline more auditable, privacy-compatible, and robust. 3.The proposed HMMIL is well aligned with the characteristics of archival multimodal data, effectively addressing the challenges of variable-length studies and missing modalities.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The description of MAGOS in the Method section is not sufficiently clear. For example, the definition of the clinical prior knowledge and the specific procedure used to obtain modality labels are not described in enough detail, and these implementation steps are not clearly illustrated in Fig. 1.The presentation could be strengthened by adding a more explicit formulation, for example with a few representative equations. 2.The evidence for clinical generalization is limited, all predictive results come from a single retrospective cohort from one hospital. In addition, the manuscript does not specify whether the dataset is public or private. If it is a public dataset, the source should be clearly stated. If it is a private dataset, the data collection process should be described in detail. 3.The authors describe the dataset only in Section 2.1, but the final classification task remains insufficiently clear. For example, it is not explicitly stated whether the goal is to distinguish patients with chronic allograft injury from normal controls. The manuscript should provide a clearer definition of the final prediction task and report performance metrics for each class. 4.Fig. 1 is too small and lacks sufficient clarity. 5.The best results in the table should be highlighted, and the table should be formatted to avoid spanning multiple pages.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper demonstrates a certain degree of novelty, but many important methodological details are not described clearly. The overall framework figure is rather coarse and lacks readability. In addition, the experimental setup—particularly the definition of the final prediction task—is not presented clearly, and the results are not reported in a sufficiently effective manner. Further revision is therefore needed.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    Many details in the manuscript are not sufficiently explained, such as the specific procedures in the methodology section, as well as the experimental settings and objectives. This lack of clarity reduces the overall readability of the paper.



Review #3

  • Please describe the contribution of the paper

    In this work the authors propose an archive-to-prediction framework, named FARSIGHT. The framework includes magos, for modality routing and ROI extraction, to handle missing modalities and automate ROI selection, whose output is then fed to HMMIL, a hierarchical multi-modal MIL, for classification. The framework is evaluated on ultrasound data, with various modalities and varying number of frames and used to classify patients of renal transplant chronic allograft injury.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Handling of real-world-data: The proposed framework, applied here to ultrasound data, handles classic real work noisy data, namely missing modalities and varying number of frames. The authors also mentioned that it has been designed to be compatible with PACS workflows. While only evaluated on data collected from one hospital, the authors outline this limitation and discuss their intention to extend to multiple centers and open source data in the future. 2.Archive-to-prediction: Though “archive” here can be slightly misleading, this paper handles an end-to-end workflow, including data curation to prepare the data for model as well as model prediction, which can save significant time in clinical workflows. Moreover, the data-curation part is designed without requiring any model training so could be applied to other cases.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Clarity: The evaluation of the individual parts of the framework and the ways in which the data is split for evaluating these is rather confusing, e.g. authors state 686 samples in test set (section 2.1) for evaluating model, 200 images, Independent test set A, randomly sampled for modality routing and region extraction (section 3.2), 400 images, internal development set A for rule refinement and pipeline verification (section 2.2). Generalizability: Though this work is presented as a general framework it is only applied to the task of renal transplant. It would have been nice to see it’s applicability to other tasks and modalities.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • In Section 2.1 a clarification is needed whether the test split is the same as the independent test set B stated later in the manuscript
    • Legends to Fig.2 and F1-score to y-axis (currently left panel not reflecting real metric based on figure title) should be included
    • I would like to see the modality counts per split (train, val, test) to ensure test set performance accurately reflects the missing modality case
    • In section 2.2 the role of the agent could be clarified
    • In section 2.2.statement “OpenCV-driven module using predefined spatial heuristics” could be better explained; what does this refer to, maybe with an example
    • In section 2.2 citation for ShuffleNet is required
    • In section 3.3 it is currently unclear how the comparison to baselines is performed - were all models retrained and tested using same dataset and splits?
    • In tables 1 and 2 the standard deviation of the reported metric should be included (also training is only nice-to-have) or statistical significance tests to back results
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors propose a method which tackles real-world hospital noisy data, such as missing-modalities in ultrasound imaging and varying number of frames per modality and automatically performs data curation with good results on a large dataset. Moreover, performance of the proposed model appears to give good classification performance (though significance of these metrics should be included) against baselines.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    Though from the authors rebuttal it is not clear whether all the reviewers comments will be sufficiently addressed in the camera-ready-version, I still believe this work is a valuable contribution worth accepting.



Author Feedback

We sincerely thank the Area Chair and Reviewers #1–#3 for their evaluation and comments. We appreciate their recognition of the novelty of our framework and its close alignment with real-world clinical practice, including archive-to-prediction workflow, handling of variable frame counts/missing modalities. Generalization/data validation. Although single-center, our cohort spans 2008–2025 and includes 4,576 patients/41,203 images from varied US devices and heterogeneous UIs. Meanwhile, external cohorts are being collected, and FARSIGHT is also used in echocardiography/breast imaging. Per MICCAI rules, we do not report these studies, but will conduct external validation and broader evaluation in future journal work. Task, dataset, and splits. We clarify that the private cohort consists of post-transplant US exams from Hospital A, with CAI status from clinical records. The task is patient-level classification of CAI vs. non-CAI patients; the latter were termed “normal controls” in Sec. 2.1.The 3,204/686/686 split is the main patient-level train/val/test split for HMMIL, with no patient overlap. Independent Sets A/B were only for MAGOS validation: A (400 images) for routing and B (200 images) for ROI extraction. They were not used to optimize MAGOS or train/test HMMIL. MAGOS/Fig.1.Fig.1 was compressed to retain content, reducing readability. Four modalities were found during data cleaning via VLM-assisted inspection of distinct US patterns; Sec. 2.1 gives counts. “Clinical prior knowledge” in Sec. 2.2 refers to radiologists’ visual knowledge for modality recognition, e.g., waveform panels (S), grayscale fields (B), red/blue overlays (C), and power-flow overlays (P), encoded as visual states/rules. In Fig.1, R1 checks inputs, R2 parses US cues, and R3 assigns S/B/C/P labels. For extraction, E1–E4 parse targets, propose/refine ROIs with modality-specific heuristics, and output ROIs. The validation/fallback block accepts/rejects/refines outputs. Due to page limits, the method was concise. If accepted, we will add equations, pseudocode, and details in the camera-ready and code repository. Experiment/modality choice. All modality rows in Table 2 use full HMMIL with only inputs changed; only “w/o” rows remove modules. Thus the full-module B+C+S result requested by R1 was reported. S alone is weak, and adding S to B+C does not improve primary test AUC or validation performance. This matches Sec.3.4, where B-mode and Color Doppler receive higher attention than S. Although S is clinically important, its diagnostic cues were not well captured by our visual modeling, so B+C was selected. All Table 1 methods used the same split/protocol. We acknowledge requests for more baselines, class-wise metrics, CIs/statistical tests, and external validation. We computed 95% CIs but removed them due to space; per MICCAI rules, we do not report extra numerical results and will consult the AC on whether they can be restored. B-mode ROI/HMMIL. The 0.80 IoU criterion suits localized Doppler/waveform panels. For B-mode, renal parenchyma is large, so fine segmentation is less critical. Preliminary trials found similar classification results between fine delineation and UI-removed fields; due to space limits, these were omitted. Architecturally, MHIM-MIL mines hard instances within a single MIL stream, whereas HMMIL embeds Top-K supervision into modality-aware hierarchical fusion. HMMIL uses modality embeddings, intra-modality Transformers, learned instance-weight pooling into modality tokens, missing-modality tokens, and inter-modality Transformer fusion for bag prediction. The Top-K loss supervises a separate instance-logit branch, reinforcing sparse positive evidence and suppressing high-scoring false positives in negative bags. We will correct Fig.2/tables/citation and incorporate all constructive suggestions. We again thank the AC/reviewers for their strong recognition.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper presents an end-to-end framework for retrospective clinical archive analysis, combining zero-shot data curation (Magos) with hierarchical multimodal MIL for patient-level prediction. Reviewers appreciated the practical motivation of the work, especially its focus on real-world archival challenges such as noisy metadata, variable numbers of images, redundant frames, and missing modalities. The privacy-preserving and auditable design of the curation pipeline was viewed positively, and the proposed HMMIL framework was considered well aligned with the characteristics of such multimodal clinical data. The end-to-end archive-to-prediction perspective is also a notable strength with clear translational relevance.

    At the same time, the reviewers raised several important concerns that need to be clarified. The main issues relate to methodological clarity and experimental support. In particular, the description of Magos remains insufficiently detailed, including the definition of the clinical prior knowledge, the routing procedure, and the implementation details illustrated in Fig. 1.The final prediction task and dataset construction are also not always clearly presented, and the different data splits used across curation and prediction experiments are somewhat confusing. On the experimental side, reviewers noted the lack of external or multi-center validation, limited comparison with multimodal baselines, unclear justification for the final modality selection, and the absence of uncertainty estimates or statistical significance testing. Questions were also raised regarding the novelty of the hard negative loss relative to prior MIL work, and regarding whether the current evidence is sufficient to support the broader generalization claims.

    Overall, the paper addresses a practically important problem and contains several promising ideas, but the current reviews indicate that key details and experimental justifications remain insufficiently resolved. I therefore encourage the authors to use the rebuttal to clarify the method description, dataset/task definition, evaluation protocol, and the rationale behind the main design and comparison choices.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper presents an end-to-end framework for retrospective clinical archive analysis combining zero-shot curation and hierarchical multimodal MIL. The problem is clinically important, the archive-to-prediction perspective is practically relevant, and the reviewers recognized the value of the contribution.

    After rebuttal, the most critical experimental concerns were resolved, and two reviewers explicitly recommend acceptance. Remaining issues regarding methodological clarity, multimodal MIL baselines, and statistical reporting are addressable and do not outweigh the paper’s merits.

    For the camera-ready version, the authors must reinstate confidence intervals and statistical tests, significantly improve the clarity of the Magos pipeline and experimental setup.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Two accept and one reject after rebuttal.

    I read through the whole contents, and think this should be above the threshold. Please address the limitaitons after.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This is a borderline paper with mixed reviews. Reviewers agreed that the paper addresses a clinically important and under-explored problem of curating large-scale, uncurated ultrasound archives for patient-level prediction, and that the proposed Magos curation pipeline is privacy-preserving, auditable, and deployable without any labeled training data. However, they also raised serious concerns about the paper’s soundness and significance: (1)​the evaluation is limited to a single private cohort with no external or multi-center validation, leaving generalizability uncertain; (2)​key design choices—such as modality selection and the justification for coarse B-mode segmentation—lack rigorous experimental support, with ablation inconsistencies only partially resolved; and (3)​the comparative evaluation omits relevant multimodal MIL baselines, making it difficult to isolate the contribution of the hierarchical design, while statistical reporting remains incomplete. The rebuttal clarified several methodological and experimental points, but did not fully resolve these concerns. After careful deliberation, I recommend accepting this manuscript, as the core contribution—an end-to-end, zero-fine-tuning framework for transforming noisy clinical archives into analyzable cohorts—represents a practical and timely advance for real-world medical AI deployment.



back to top